The two conclusions of this study are divergent. First, a genetic difference exists between those who have the most severe course of COVID-19 and the general population. Second, we were not able to exploit this difference to develop a clinically useful test to distinguish between people who will experience a severe course of the disease and those who will not. We could only demonstrate a genetic risk test with an AUC of 0.51, just slightly above 0.50 which represents random guessing.
Although the AUC we found here is too low to be clinically useful, several avenues for improving the AUC exist. We were constrained by the data available to compare those who had a severe reaction to COVID-19 with the general population, but the general population probably contains a substantial number of people who would also have a severe reaction to COVID-19. A better approach would be to compare those who had a severe reaction to COVID-19 with those who were asymptomatic or had a mild reaction. Simply having a much larger number of patients who had a severe reaction might also lead to an increase in AUC.
Changes in our feature selection and classification algorithm might also improve the AUC. Our feature selection algorithm that transformed “l2r” data into our final chromosomal-scale length variation data took averages over each quarter of a chromosome. We could instead include smaller chromosome segments. Generally, we need the number of features to be much less than the number of observations (patients). So, an increase in the number of observations would allow an increase in the number of features. Also, an alternative machine learning algorithm might improve the AUC. Different algorithms perform differently on different classes of problems and XGBoost generally performs well on tabular data. We did a brief test of different algorithms before choosing XGBoost as the best solution for this problem. But, for instance, a deep learning algorithm might have better performance with proper tuning.
Our results add to the recent work done by others on the link between genetics and severity of COVID-19. For instance, one study from the Netherlands identified four young men from two different families who had severe symptoms of COVID-19 and no preexisting medical conditions. Detailed genetic studies revealed that these four men all had a rare loss of function variant of TLR7, which lies on the X-chromosome.
A detailed study of this UK Biobank COVID-19 dataset found that Black and Asian patients were at a significantly higher risk of testing positive compared to white patients . This study also attempted to derive a polygenic risk score. However, when they applied the polygenic risk score to a hold-out group, they found that the mean score was indistinguishable between the group of people who had tested positive and the group that had no positive test. In comparison, our work found that these two groups are distinguishable with a genetic risk score, but only very slightly. We measured the AUC at 0.51. They  do not report an AUC, but an indistinguishable test is the equivalent of an AUC of 0.50.
Other more comprehensive metastudies have identified one specific genetic component behind the severity of COVID-19. For instance, one study of COVID-19 patients who experienced respiratory failure at seven hospitals in Italy and Spain found a fairly strong association in a cluster of genes lying on part of chromosome 3 and a borderline association in chromosome 9 encompassing the ABO blood group locus . The “ANA_B2” June 2020 results posted by the COVID-19 Host Genetics Initiative [12,13], also indicate a strong association in Chromosome 3, but fail to reproduce the association in chromosome 9. The COVID-19 Host Genetics Initiative “ANA_B2” study compares hospitalized COVID-19 patients to the general population and are mostly derived from patients in Europe and Brazil. Neither study attempted to derive a genetic risk score.
This study has several weaknesses. First, we cannot attribute the severity of COVID-19 to particular genetic variants. This study only finds correlations and does not establish a cause and effect. Second, while it is possible that these correlations relate to underlying biology, it is also possible that the correlations are related to ancestral differences that translate to socio-economic differences. COVID-19 severity is known to be correlated with racial/ethnic background[14,15]. The small effect that we measured might be simply due to the larger complex effect of racial/ethnic disparities in COVID-19 severity.